Dynamic management of scratchpad memory based on compiler driven approach

Scratchpad memory (SPM) is commonly used in embedded system because it is more efficient than cache in both performance and energy, and it can satisfy real time requirements as well. A compiler driven method is proposed in this paper to optimally mapping code segments to SPM at runtime. The algorithm uses compiler to identify hotspots in programs based on a greedy cost model, and utilizes DMA transfer instead of explicit load and store operations to fill SPM. Compared to cache method, experimental results show performance improvements by 27.3% and energy savings by 35.5%.